Vinogradov’s estimate for exponential sums over primes

Jordan Bell
February 23, 2016

1 Introduction

For x∈ℝ, let [x] be the greatest integer ≤x, let R⁢(x)=x-[x], and let

∥x∥=min⁡{R⁢(x),1-R⁢(x)}=minm∈ℤ⁡|x-m|.

In this note I work through Chapters 24 and 25 of Harold Davenport, Multiplicative Number Theory, third ed.11 1 Many of the manipulations of sums in these chapters are hard to follow, and I greatly expand on the calculations in Davenport. The organization of the proof in Davenport seems to be due to Vaughan. I have also used lecture notes by Andreas Strömbergsson, http://www2.math.uu.se/~astrombe/analtalt08/www_notes.pdf, pp. 245–257. Another set of notes, which I have not used, are http://jonismathnotes.blogspot.ca/2014/11/prime-exponential-sums-and-vaughans.html

We end up proving that there is some constant C such that if α∈ℝ and |α-aq|≤1q2, with a positive and gcd⁡(a,q)=1, then for any N≥2,

2 The von Mangoldt function

Let Λ be the von Mangoldt function: Λ⁢(n)=log⁡p if n=pα for some prime p and α≥1, and Λ⁢(n)=0 otherwise. For example, Λ⁢(4)=log⁡2, Λ⁢(11)=log⁡11, and Λ⁢(12)=0. This satisfies

log⁡n=∑d∣nΛ⁢(d),

and so by the Möbius inversion formula,

Λ⁢(n)=∑d∣nμ⁢(n/d)⁢log⁡d.

For n>1 it is a fact that22 2 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 235, Theorem 263.

∑d∣nμ⁢(d)=0.

Write ψ⁢(x)=∑n≤xΛ⁢(n). It is a fact that ψ⁢(x)=O⁢(x).33 3 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 341, Theorem 414. (The prime number theorem states ψ⁢(x)∼x.)

The derivative of the Riemann zeta function is

ζ′⁢(s)=-∑n=1∞n-s⁢log⁡n,Re⁢s>1.

The Euler product for the Riemann zeta function is

ζ⁢(s)=∑m=1∞m-s=∏p(1-p-s)-1,Re⁢s>1.

Then

-log⁡ζ⁢(s)=∑plog⁡(1-p-s),

so, for Re⁢s>1,

-ζ′⁢(s)ζ⁢(s) =∑pp-s⁢log⁡p1-p-s
=∑plog⁡p⁢∑l=1∞p-l⁢s
=∑p∑l=1∞Λ⁢(pl)⁢(pl)-s
=∑n=1∞Λ⁢(n)⁢n-s.

Let U,V≥2, U⁢V≤N, and write

FU⁢(s)=∑j≤UΛ⁢(j)⁢j-s,GV⁢(s)=∑k≤Vμ⁢(k)⁢k-s.

For Re⁢s>1,

-ζ′⁢(s)ζ⁢(s)=FU⁢(s)-ζ⁢(s)⁢FU⁢(s)⁢GV⁢(s)-ζ′⁢(s)⁢GV⁢(s)+(-ζ′⁢(s)ζ⁢(s)-FU⁢(s))⁢(1-ζ⁢(s)⁢GV⁢(s)).

First,44 4 Harold Davenport, Multiplicative Number Theory, third ed., p. 138, Chapter 24.

FU⁢(s)=∑n=1∞a1⁢(n)⁢n-s

for

a1⁢(n)={Λ⁢(n)n≤Ua1⁢(n)=0n>U.

Second,

-ζ⁢(s)⁢FU⁢(s)⁢GV⁢(s)=∑n=1∞a2⁢(n)⁢n-s

for

a2⁢(n)=-∑m⁢j⁢k=n,m≥1,j≤U,k≤V1⋅Λ⁢(j)⋅μ⁢(k)=-∑d∣n∑d⁢j⁢k=n,j≤U,k≤VΛ⁢(j)⁢μ⁢(k).

Third,

-ζ′⁢(s)⁢GV⁢(s)=∑n=1∞a3⁢(n)⁢n-s

for

a3⁢(n)=∑m⁢k=n,m≥1,k≤Vlog⁡(m)⋅μ⁢(k)=∑d∣nlog⁡d⁢∑d⁢k=n,k≤Vμ⁢(k).

Fourth,

-ζ′⁢(s)ζ⁢(s)-FU⁢(s)=∑m>UΛ⁢(m)⁢m-s;

and

ζ⁢(s)⁢GV⁢(s)=∑n=1∞(∑m⁢k=n,m≥1,k≤V1⋅μ⁢(k))⁢n-s=∑n=1∞(∑d∣n,d≤Vμ⁢(d))⁢n-s

whence

1-ζ⁢(s)⁢GV⁢(s)=-∑h=2∞(∑d∣h,d≤Vμ⁢(d))⁢h-s;

thus

(-ζ′⁢(s)ζ⁢(s)-FU⁢(s))⁢(1-ζ⁢(s)⁢GV⁢(s))=∑n=1∞a4⁢(n)⁢n-s

for

a4⁢(n)=-∑m⁢h=n,m>U,h>1Λ⁢(m)⁢(∑d∣h,d≤Vμ⁢(d)).

We have

Λ⁢(n)=a1⁢(n)+a2⁢(n)+a3⁢(n)+a4⁢(n).

3 Sums involving the von Mangoldt function

Let f be an arithmetical function with |f|≤1 and write

Si=∑n≤Nf⁢(n)⁢ai⁢(n),

for which

∑n≤Nf⁢(n)⁢Λ⁢(n)=S1+S2+S3+S4.
S1 =∑n≤Uf⁢(n)⁢Λ⁢(n)
S2 =-∑n≤Nf⁢(n)⁢∑d∣n∑d⁢j⁢k=n,j≤U,k≤VΛ⁢(j)⁢μ⁢(k)
S3 =∑n≤Nf⁢(n)⁢∑d∣nlog⁡d⁢∑d⁢k=n,k≤Vμ⁢(k)
S4 =-∑n≤Nf⁢(n)⁢∑m⁢h=n,m>U,h>1Λ⁢(m)⁢∑d∣h,d≤Vμ⁢(d).
Lemma 1.

|S1|=O⁢(U).

Proof.

As |f|≤1,

|S1|≤∑n≤UΛ⁢(n)=ψ⁢(U)=O⁢(U).

∎

Lemma 2.
|S2|≤log⁡U⁢V⋅∑h≤U⁢V|∑r≤N/hf⁢(r⁢h)|.
Proof.
S2 =-∑n≤Nf⁢(n)⁢∑d∣n∑d⁢j⁢k=n,j≤U,k≤VΛ⁢(j)⁢μ⁢(k)
=-∑h≤U⁢V(∑j⁢k=h,j≤U,k≤VΛ⁢(j)⁢μ⁢(k))⁢∑r≤N/hf⁢(r⁢h).

For h≤U⁢V, ∑j∣hΛ⁢(j)=log⁡h≤log⁡U⁢V, so

|S2|≤log⁡U⁢V⋅∑h≤U⁢V|∑r≤N/hf⁢(r⁢h)|.

∎

Lemma 3.
|S3|≤log⁡N⋅∑k≤Vmax1≤w≤N/k⁡|∑w≤h≤N/kf⁢(k⁢h)|.
Proof.
S3 =∑n≤Nf⁢(n)⁢∑d∣nlog⁡d⁢∑d⁢k=n,k≤Vμ⁢(k)
=∑k≤Vμ⁢(k)⁢∑h≤N/kf⁢(k⁢h)⁢log⁡h
=∑k≤Vμ⁢(k)⁢∑h≤N/kf⁢(k⁢h)⁢∫1hd⁢ww
=∑k≤Vμ⁢(k)⁢∫1N/k∑w≤h≤N/kf⁢(k⁢h)⁢d⁢ww
=∫1N∑k≤Vμ⁢(k)⁢∑w≤h≤N/kf⁢(k⁢h)⁢d⁢ww.

Then

|S3| ≤max1≤w≤N⁡|∑k≤Vμ⁢(k)⁢∑w≤h≤N/kf⁢(k⁢h)|⋅∫1Nd⁢ww
≤log⁡N⋅∑k≤Vmax1≤w≤N/k⁡|∑w≤h≤N/kf⁢(k⁢h)|.

∎

Lemma 4.
|S4|≪N1/2⁢(log⁡N)3⁢maxU≤M≤N/V⁡Δ,

for

Δ=maxV<j≤N/M(∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf(mj)f⁢(m⁢k)¯|)1/2.
Proof.

For bm,ck∈ℂ, using the Cauchy-Schwarz inequality,

|∑M<m≤2⁢Mbm⁢∑V<k≤N/mck⁢f⁢(m⁢k)|≤(∑M≤m≤2⁢M|bm|2)1/2⁢(∑M≤m≤2⁢M|∑V<k≤N/mck⁢f⁢(m⁢k)|2)1/2,

and

∑M≤m≤2⁢M|∑V<k≤N/mck⁢f⁢(m⁢k)|2=∑M≤m≤2⁢M(∑V<j≤N/mcj⁢f⁢(m⁢j))⁢(∑V<k≤N/mck⁢f⁢(m⁢k)¯)=∑V<j≤N/M∑V<k≤N/Mcj⁢ck¯⁢∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯,

so, using |cj⁢ck|≤12⁢|cj|2+12⁢|ck|2,

∑M≤m≤2⁢M|∑V<k≤N/mck⁢f⁢(m⁢k)|2≤∑V<j≤N/M∑V<k≤N/M(12⁢|cj|2+12⁢|ck|2)⁢|∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯|=∑V<j≤N/M|cj|2⁢∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯|≤(∑V<j≤N/M|cj|2)⁢maxV<j≤N/M⁢∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯|.

Therefore,

|∑M<m≤2⁢Mbm⁢∑V<k≤N/mck⁢f⁢(m⁢k)|≤Δ⁢(∑M≤m≤2⁢M|bm|2)1/2⁢(∑j≤N/M|cj|2)1/2

for

Δ=(maxV<j≤N/M⁢∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯|)1/2.

Because ∑d∣hμ⁢(d)=0 for h>1, if 1<h≤V then ∑d∣h,d≤Vμ⁢(d)=0. Thus

S4 =-∑n≤Nf⁢(n)⁢∑m⁢h=n,m>U,h>1Λ⁢(m)⁢∑d∣h,d≤Vμ⁢(d)
=-∑U<m<N/VΛ⁢(m)⁢∑V<k≤N/mf⁢(m⁢k)⁢∑d∣k,d≤Vμ⁢(d)
=-∑U<m≤N/VΛ⁢(m)⁢∑V<k≤N/mf⁢(m⁢k)⁢∑d∣k,d≤Vμ⁢(d)
=-∑M∈{U,2⁢U,4⁢U,…},M<N/V∑M<m≤min⁡(N/V,2⁢M)Λ⁢(m)⁢∑V<k≤N/mf⁢(m⁢k)⁢∑d∣k,d≤Vμ⁢(d),

so

|S4|≤(log2⁡NU⁢V)⁢maxU≤M≤N/V⁡|∑M<m≤min⁡(N/V,2⁢M)Λ⁢(m)⁢∑V<k≤N/mf⁢(m⁢k)⁢∑d∣k,d≤Vμ⁢(d)|.

Define bm=Λ⁢(m) for m≤N/V and bm=0 for m>N/V, and ck=∑d∣k,d≤Vμ⁢(d). Using the above we get

|S4| ≪(log⁡N)⁢maxU≤M≤N/V⁡|∑M<m≤2⁢Mbm⁢∑V<k≤N/mck⁢f⁢(m⁢k)|
≪(log⁡N)⁢maxU≤M≤N/V⁡Δ⁢(∑M≤m≤2⁢M|bm|2)1/2⁢(∑j≤N/M|cj|2)1/2
≪(log⁡N)⁢maxU≤M≤N/V⁡Δ⁢(∑M≤m≤2⁢MΛ⁢(m)2)1/2⁢(∑j≤N/Md⁢(k)2)1/2

On the one hand,

∑m≤yΛ⁢(m)2≤(log⁡y)⁢∑m≤yΛ⁢(m)=O⁢(y⁢log⁡y).

On the other hand, let h be the multiplicative arithmetic function such that for prime p and for nonnegative integer a, h⁢(pa)=2⁢a+1. The divisor function satisfies d⁢(pa)=a+1, and

∑d∣pah⁢(d) =∑0≤b≤ah⁢(pb)
=∑0≤b≤a(2⁢b+1)
=a+1+2⁢∑0≤b≤ab
=a+1+2⋅a⁢(a+1)2
=a+1+a2+a
=a2+2⁢a+1
=d⁢(pa)2.

Hence, as d↦h⁢(d)d is multiplicative and nonnegative,

∑k≤yd⁢(k)2 =∑k≤y∑d∣kh⁢(d)
=∑d≤yh⁢(d)⁢∑k⁢d≤y1
=∑d≤yh⁢(d)⋅[y/d]
≤y⁢∑d≤yh⁢(d)d
≤y⁢∏p≤y∑a=0∞h⁢(pa)⁢p-a
=y⁢∏p≤y∑a=0∞(2⁢a+1)⁢p-a.

But, for 0<x<1,

(1-x)-3=(∑a=0∞xa)3=∑a=0∞12⁢(a+1)⁢(a+2)⁢xa≥∑a=0∞(2⁢a+1)⁢xa,

so

∑k≤yd⁢(k)2≤y⁢(∑p≤y(1-p-1)-1)3.

Merten’s theorem55 5 G. H. Hardy and E. M. Wright, An Introduction to the Theory of Numbers, fifth ed., p. 351, Theorem 429. tells us

∏p≤y(1-1p)∼e-γlog⁡y,

where γ is Euler’s constant, and using this,

∑k≤yd⁢(k)2=O⁢(y⁢(log⁡y)3).

We have therefore got for U≤M≤N/V,

(∑M≤m≤2⁢MΛ⁢(m)2)1/2⁢(∑k≤N/Md⁢(k)2)1/2≤(∑m≤2⁢MΛ⁢(m)2)1/2⁢(∑k≤N/Md⁢(k)2)1/2≤(O(MlogM))1/2(O(N/M(log(N/M))3)1/2=O⁢(N1/2⁢(log⁡N)2).

What we now have is

|S4| ≪(log⁡N)⋅N1/2⁢(log⁡N)2⋅maxU≤M≤N/V⁡Δ,

proving the claim. ∎

Putting together the estimates for S1,S2,S3,S4 gives, for |f|≤1, and U,V≥2, U⁢V≤N,

∑n≤Nf⁢(n)⁢Λ⁢(n) ≪U+(log⁡N)⁢∑h≤U⁢V|∑r≤N/hf⁢(r⁢h)|
+(log⁡N)⁢∑k≤Vmax1≤w≤N/k⁡|∑w≤h≤N/kf⁢(k⁢h)|
+N1/2⁢(log⁡N)3⁢maxU≤M≤N/V⁡Δ,

for

Δ=maxV<j≤N/M(∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf(mj)f⁢(m⁢k)¯|)1/2.

4 Exponential sums

For β∈ℝ, on the one hand

|∑N1≤n≤N2e2⁢π⁢i⁢β⁢n|≤N2-N1+1.

On the other hand,

∑N1≤n≤N2e2⁢π⁢i⁢β⁢n=∑0≤n≤N2-N1e2⁢π⁢i⁢β⁢n=1-e2⁢π⁢i⁢β⁢(N2-N1+1)1-e2⁢π⁢i⁢β

and hence

|∑N1≤n≤N2e2⁢π⁢i⁢β⁢n|≤2|1-e2⁢π⁢i⁢β|=1|sin⁡π⁢β|≤12⁢∥β∥.

Thus

|∑N1≤n≤N2e2⁢π⁢i⁢β⁢n|≪min⁡{N2-N1,1∥β∥}.

Let α∈ℝ and let f⁢(n)=e2⁢π⁢i⁢α⁢n. Then

∑h≤U⁢V|∑r≤N/hf⁢(r⁢h)| ≤∑h≤U⁢Vmin⁡{Nh,1∥h⁢α∥}

and

∑k≤Vmaxw⁡|∑w≤h≤N/kf⁢(r⁢t)| ≪∑k≤Vmaxw⁡min⁡{Nk-w,1∥k⁢α∥}
≪∑k≤Vmin⁡{Nk,1∥k⁢α∥}.

Let SN⁢(α)=∑n≤NΛ⁢(n)⁢f⁢(n). By what we have worked out,

|SN⁢(α)| ≪U+(log⁡N)⁢∑h≤U⁢Vmin⁡{Nh,1∥h⁢α∥}
+(log⁡N)⁢∑k≤Vmin⁡{Nk,1∥k⁢α∥}
+N1/2⁢(log⁡N)3⁢maxU≤M≤N/V⁡Δ,

for

Δ=maxV<j≤N/M(∑V<k≤N/M|∑M≤m≤2⁢M,m≤N/j,m≤N/kf(mj)f⁢(m⁢k)¯|)1/2.

We calculate

∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯ =∑M≤m≤2⁢M,m≤N/j,m≤N/ke2⁢π⁢i⁢α⁢m⁢(j-k)

so

|∑M≤m≤2⁢M,m≤N/j,m≤N/kf⁢(m⁢j)⁢f⁢(m⁢k)¯|≪min⁡{M,1∥(j-k)⁢α∥}.

We now have

|SN⁢(α)| ≪U+(log⁡N)⁢∑h≤U⁢Vmin⁡{Nh,1∥h⁢α∥}
+(log⁡N)⁢∑k≤Vmin⁡{Nk,1∥k⁢α∥}
+N1/2(logN)3maxU≤M≤N/VmaxV<j≤N/M(∑V<k≤N/Mmin{M,1∥(k-j)⁢α∥})1/2.

But for V<j≤N/M, a fortiori 0≤j≤N/M, whence

∑V<k≤N/Mmin⁡{M,1∥(k-j)⁢α∥} ≤∑0≤k≤N/Mmin⁡{M,1∥(k-j)⁢α∥}
≤∑|m|≤N/Mmin⁡{M,1∥m⁢α∥}
=M+2⁢∑1≤m≤N/Mmin⁡{M,1∥m⁢α∥}
≪M+∑1≤m≤N/Mmin⁡{Nm,1∥m⁢α∥}.

Summarizing, we have the following.

Theorem 5.

For α∈R and U,V≥2,U⁢V≤N,

|SN⁢(α)| ≪U+(log⁡N)⁢∑h≤U⁢Vmin⁡{Nh,1∥h⁢α∥}
+N1/2(logN)3maxU≤M≤N/V(M+∑1≤m≤N/Mmin{Nm,1∥m⁢α∥})1/2.

5 Diophantine approximation

Theorem 6.

There is some C such that for all 0<α<1, if |α-aq|≤1q2, gcd⁡(a,q)=1, and T≥1, then

∑t≤Tmin⁡{Nt,1∥t⁢α∥}≤C⁢(Nq+T+q)⁢log⁡(2⁢q⁢T).
Proof.

Write β=α-aq. Then

∑t≤Tmin⁡{Nt,1∥t⁢α∥} ≤∑0≤h≤T/q∑1≤r≤qmin⁡{Nh⁢q+r,1∥h⁢q⁢α+r⁢α∥}
=∑0≤h≤T/q∑1≤r≤qmin⁡{Nh⁢q+r,1∥r⁢aq+h⁢q⁢β+r⁢β∥}.

If h=0 and 1≤r≤q2, then, using ∥x-y∥≥∥x∥-∥y∥ and |β|≤1q2,

1∥r⁢aq+h⁢q⁢β+r⁢β∥=1∥r⁢aq+r⁢β∥≤1∥r⁢aq∥-∥r⁢β∥≤1∥r⁢aq∥-12⁢q,

and, as gcd⁡(a,q)=1,

∑1≤r≤q21∥r⁢aq∥-12⁢q ≤∑1≤m<q1∥mq∥-12⁢q
≤2⁢∑1≤m≤q21mq-12⁢q
=∑1≤m≤q24⁢q2⁢m-1
≤4⁢q⁢∑1≤m≤q-11m
≤4⁢q⁢log⁡(2⁢q).

Otherwise, 1≤h≤T/q or q2<r≤q, and then h⁢q+r≥12⁢(h+1)⁢q, and the sum over these indices is

≪∑0≤h≤T/q∑1≤r≤qmin⁡{N(h+1)⁢q,1∥r⁢aq+h⁢q⁢β+r⁢β∥}.

So we have got

∑t≤Tmin⁡{Nt,1∥t⁢α∥}≪q⁢log⁡(2⁢q)+∑0≤h≤T/q∑1≤r≤qmin⁡{N(h+1)⁢q,1∥r⁢aq+h⁢q⁢β+r⁢β∥}.

Let 1≤h≤T/q, let I=[A,B] be a closed arc in ℝ/ℤ of measure q-1, and let J=[A-h⁢q⁢β-q-1,B-h⁢q⁢β+q-1]⊂ℝ/ℤ. For 1≤r≤q, if r⁢aq+h⁢q⁢β+r⁢β∈I then r⁢aq+r⁢β∈[A-h⁢q⁢β,B-h⁢q⁢β], and as |r⁢β|≤q-1, then r⁢aq∈J. As J is a closed arc with measure 3⁢q-1 and aq,2⁢aq,…,q⋅aq are distinct in ℝ/ℤ, due to gcd⁡(a,q)=1, there are at most four r, 1≤r≤q, for which r⁢aq∈J. Therefore there are at most four r, 1≤r≤q, for which r⁢aq+h⁢q⁢β+r⁢β∈I.

For 0≤j≤q-1, let Ij=[j⁢q-1,(j+1)⁢q-1]. If r⁢aq+h⁢q⁢β+r⁢β∈Ij⊂ℝ/ℤ, then

∥r⁢aq+h⁢q⁢β+r⁢β∥≥min⁡{j⁢q-1,1-(j+1)⁢q-1}

i.e.

1∥r⁢aq+h⁢q⁢β+r⁢β∥≤qmin⁡{j,q-j-1}.

Therefore, for 1≤r≤q with r⁢aq+h⁢q⁢β+r⁢β∈Ij⊂ℝ/ℤ,

min⁡{N(h+1)⁢q,1∥r⁢aq+h⁢q⁢β+r⁢β∥} ≤min⁡{N(h+1)⁢q,qmin⁡{j,q-j-1}}
≤{N(h+1)⁢qj=0,q-1qmin⁡{j,q-j-1}1≤j≤q-2.

We have just established that for each 0≤j≤q-1 there are at most four 1≤r≤q such that r⁢aq+h⁢q⁢β+r⁢β∈Ij⊂ℝ/ℤ, and hence

∑0≤h≤T/q∑1≤r≤qmin⁡{N(h+1)⁢q,1∥r⁢aq+h⁢q⁢β+r⁢β∥}≤∑0≤h≤T/q∑0≤j≤q-14⋅{N(h+1)⁢qj=0,q-1qmin⁡{j,q-j-1}1≤j≤q-2=∑0≤h≤T/q(8⁢N(h+1)⁢q+∑1≤j≤q-2qmin⁡{j,q-j-1})≤8⁢Nq⁢∑1≤h≤Tq+11h+1+(Tq+1)⋅2⁢q⁢∑1≤j≤q/21j≪Nq⁢log⁡(2⁢T/q)+(Tq+1)⋅q⁢log⁡q.

Putting things together,

∑t≤Tmin⁡{Nt,1∥t⁢α∥} ≪q⁢log⁡(2⁢q)+Nq⁢log⁡(2⁢T)+(Tq+1)⁢q⁢log⁡(2⁢q)
≪q⁢log⁡(2⁢q⁢T)+Nq⁢log⁡(2⁢q⁢T)+T⁢log⁡(2⁢q⁢T).

∎

We now combine Theorem 5 and Theorem 6. For U,V≥2,U⁢V≤N,T≥1, |α-aq|≤1q2, gcd⁡(a,q)=1,

|SN⁢(α)| ≪U+(log⁡N)⁢(Nq+U⁢V+q)⁢log⁡(2⁢q⁢U⁢V)
+N1/2(logN)3maxU≤M≤N/V(M+(Nq+NM+q)log(2qN/M))1/2
≪U+(log⁡2⁢q⁢N)3⁢(Nq+U⁢V+q)
+N1/2(logqN)7/2maxU≤M≤N/V(M+Nq+NM+q)1/2
≪U+(log⁡2⁢q⁢N)3⁢(Nq+U⁢V+q)
+N1/2⁢(log⁡q⁢N)7/2⁢((U+Nq+NU+q)1/2+(NV+Nq+V+q)1/2).

Now take U=V, for which

|SN⁢(α)| ≪U+(log⁡2⁢q⁢N)3⁢(Nq+U2+q)
+N1/2⁢(log⁡q⁢N)7/2⁢(U+Nq+NU+q)1/2
≪U+(log⁡2⁢q⁢N)3⁢(Nq+U2+q)
+N1/2⁢(log⁡q⁢N)7/2⁢(U1/2+N1/2⁢q-1/2+N1/2⁢U-1/2+q1/2)
=U+(log⁡2⁢q⁢N)3⁢(Nq+U2+q)
+(log⁡q⁢N)7/2⁢(N1/2⁢U1/2+N⁢q-1/2+N⁢U-1/2+N1/2⁢q1/2).

For U=N2/5 we get the following.

Theorem 7.

There is some C such that if α∈R, |α-aq|≤1q2, a≥1, gcd⁡(a,q)=1, then for any N≥1,

|SN⁢(α)|≤C⁢(N⁢q-1/2+N4/5+N1/2⁢q1/2)⁢(log⁡N)4.